CUDA operation dispatch surface.

The submodules separate tensor views, reductions, neural-network kernels, linear algebra, FFT, and other CUDA-backed eager operations while keeping the public import path stable.